Multi-fuzzy representation and anomaly detection method for spatiotemporal dynamics of gas turbine equipment

Abstract

The application discloses a kind of multiple fuzzy representation and abnormality detection method for the spatiotemporal dynamics of gas turbine equipment.The abnormality detection model of the application realizes the effective separation and explicit extraction of multiple spatiotemporal relationships by constructing a customized cross-spatiotemporal graph to perceive the dynamics of time-dependent relationships, the continuity of spatial coupling relationships and hysteresis in real time, and then obtains fine-grained time and spatial dependence to help achieve accurate anomaly detection. By designing a fuzzy embedding method to map process data to a high-dimensional fuzzy space to give it multiple fuzzy states, and by jointly representing fuzzy states to enhance the ability to describe multiple spatiotemporal relationships and solve their uncertainty problems. The abnormality detection model of the application not only adaptively realizes the explicit separation and representation of multiple spatiotemporal relationships, but also embeds process data into fuzzy states, providing an efficient and reliable new approach for anomaly detection of gas turbine equipment operating conditions.

Description

Technical Field

[0001] This invention belongs to the field of industrial process anomaly detection, and in particular relates to a multi-spatial-temporal fuzzy characterization method that can perceive the dynamics of temporal dependencies, the continuity and lag of spatial coupling relationships of gas turbine equipment subsystems in real time. It can not only achieve effective separation and explicit extraction of multiple spatiotemporal relationships, but also solve the uncertainty problem of multiple spatiotemporal relationships, and finally obtain fine-grained temporal and spatial dependencies to achieve accurate detection of gas turbine equipment operation anomalies. Background Technology

[0002] As modern industry rapidly develops towards large-scale, high-end, complex, and intelligent operations, the safety of industrial processes faces greater challenges. The sheer size of systems and the complexity of equipment increase the risk of malfunctions, and the consequences of such malfunctions are more severe. In some large-scale industrial processes, such as gas turbine power generation and petrochemicals, malfunctions can lead to property damage, personal injury, and safety accidents. Therefore, to ensure the safe and economical operation of large-scale industrial processes, precise and meticulous anomaly detection of critical industrial equipment is essential.

[0003] Gas turbines, as internal combustion mechanical devices that utilize the combustion of natural or liquid fuels to generate high-temperature, high-pressure gas, driving bearings to rotate and produce useful work, have been widely used in power generation, aviation, aerospace, and marine engineering. Gas turbines comprise numerous subsystems and various auxiliary systems, such as the turbine, combustion chamber, and compressor, with complex spatiotemporal relationships existing between these subsystems. Specifically, each subsystem frequently experiences peak shaving in the time dimension, exhibiting significant temporal dynamic characteristics; during actual operation, subsystems may experience simultaneous or even cross-time-sequence influences, demonstrating complex spatial coupling characteristics. In practical applications, gas turbine anomalies can be categorized into abnormal deviations in the temporal correlation within subsystems and abnormal spatial coupling relationships between subsystems. These two types of anomalies have different characteristics and manifestations. Therefore, effectively and explicitly separating the multiple spatiotemporal relationships (temporal dependencies within subsystems, simultaneous spatial coupling relationships between subsystems, and cross-time-sequence spatial coupling relationships between subsystems) existing within gas turbine subsystems is a pressing problem to be solved in the task of detecting anomalies in gas turbine operation. However, current research on anomaly detection technology for gas turbine equipment operation status in China is still insufficient, and the gas turbine equipment industry is still in an immature state. Therefore, it is urgent to promote and improve research on anomaly detection for gas turbine equipment operation status. Existing methods for anomaly detection of gas turbine equipment operation status often confuse temporal and spatial information, leading to interference within multiple spatiotemporal relationships. This results in an inability to provide explicit and accurate multi-temporal relationship representations, reducing the interpretability of anomaly detection results. Furthermore, existing methods do not recognize that frequent switching of operating conditions may cause uncertainty in multiple spatiotemporal relationships, which may lead to more false alarms and false negatives, further reducing the accuracy of anomaly detection.

[0004] In summary, for actual gas turbine equipment, multiple spatiotemporal relationships exist between its subsystems. This complexity and uncertainty increase the difficulty of detecting operational anomalies. Therefore, a multi-spatiotemporal fuzzy representation method should be designed to perceive the dynamics of temporal dependencies and the continuity and lag of spatial coupling relationships in real time. This will enable the effective separation and explicit extraction of multiple spatiotemporal relationships, address their uncertainty issues, and ultimately obtain fine-grained temporal and spatial dependencies, thereby improving the accuracy and interpretability of anomaly detection results. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing anomaly detection technologies targeting gas turbine equipment, which cannot explicitly separate multiple spatiotemporal relationships, by providing a multi-fuzzy representation and anomaly detection method for the spatiotemporal dynamics of gas turbine equipment.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A multi-fuzzy representation and anomaly detection method for the spatiotemporal dynamics of gas turbine equipment is disclosed, specifically comprising: acquiring process data segments of the gas turbine equipment in real time and inputting them into a trained anomaly detection model; determining whether the gas turbine equipment is abnormal based on the difference between the reconstructed process data at the last moment output by the anomaly detection model and the true value; wherein, the anomaly detection model is trained through the following steps:

[0008] Step 1: Collect process data under the frequent changes in operating conditions of the gas turbine during normal operation, and construct samples by sliding window partitioning to obtain a training dataset; where each sample is time series data of T times for N measured variables;

[0009] Step 2: Construct a training network, which includes a context information fusion module, a fuzzy embedding data generation module, and a multi-temporal relationship explicit separation and representation module. The context information fusion module integrates the context information of the input sample along the time dimension into each measured variable based on a one-dimensional convolutional kernel to obtain a context-fused sample. The fuzzy embedding data generation module includes a fuzzy membership value generation layer, a fuzzy membership vector generation layer, and a fuzzy embedding data generation layer. The fuzzy membership value generation layer generates Q fuzzy membership values ​​for each time step and each measured variable in the context-fused sample based on Q membership functions. The fuzzy membership vector generation layer generates Q fuzzy membership values ​​for each time step and each measured variable. The corresponding Q fuzzy membership values ​​are concatenated and enhanced to obtain a fuzzy membership vector; the fuzzy embedding data generation layer is used to concatenate the fuzzy membership vectors of N measured variables at T time points to obtain fuzzy embedding data; the multiple spatiotemporal relationship explicit separation and representation module includes a spatiotemporal graph generation module, a multi-level spatiotemporal co-convolutional network, and a prediction layer, wherein the spatiotemporal graph generation module is used to generate T-1 spatiotemporal graphs based on the fuzzy embedding data; the T-1 spatiotemporal graphs and the T-1 fuzzy embedding data are corresponding one-to-one with the sub-segments recombined from the leading time dimension according to every two time points, the spatiotemporal graph is a 2N×2N matrix, and the value of the i-th row and j-th column of the spatiotemporal graph is the correlation of the vectors in the i-th and j-th rows of the corresponding sub-segments; the multi-level spatiotemporal co-convolutional network contains the T-1 spatiotemporal graphs. Figure 1 A corresponding T-1 level is defined, with each level containing L fully connected layers and max pooling layers connected sequentially. The input of the first fully connected layer in each level is the corresponding normalized spatiotemporal graph and sub-segments, and the input of the 2nd to Lth fully connected layers is the corresponding normalized spatiotemporal graph and the output of the previous fully connected layer. The max pooling layer is used to aggregate the outputs of each fully connected layer and perform max pooling. The prediction layer is used to predict the process data reconstruction result of the last moment of the sample based on the concatenation result of the output of the T-1 level of the multi-level spatiotemporal collaborative convolutional network.

[0010] Step 3: Input each sample from Step 1 into the training network and train it by minimizing the loss function until the loss function converges or the set number of training iterations is reached; the loss function includes the reconstruction error of the multiple spatiotemporal relationship explicit separation and representation module; the anomaly detection model consists of a trained context information fusion module, a fuzzy embedding data generation module, a reconstruction module, and a multiple spatiotemporal relationship explicit separation and representation module.

[0011] Furthermore, the training network also includes a reconstruction module, which comprises a fuzzy rule layer, a fuzzy conclusion layer, and a defuzzification layer. The fuzzy rule layer contains Q neurons, and each neuron uses the T-norm operator to calculate the trigger strength of the corresponding fuzzy rule based on the fuzzy membership vector output by the fuzzy membership function of the nth variable in the fuzzy embedded data. The fuzzy conclusion layer is used to normalize and concatenate the trigger strengths output by the fuzzy rule layer. The defuzzification layer is used to perform a defuzzification operation on the output of the fuzzy conclusion layer using a nonlinear mapping to generate reconstructed samples. The loss function also includes the reconstruction error of the reconstruction module.

[0012] Furthermore, the anomaly detection model is composed of a trained context information fusion module, a fuzzy embedded data generation module, and a reconstruction module, or it is composed of a trained context information fusion module, a fuzzy embedded data generation module, a multiple spatiotemporal relationship explicit separation and representation module, and a reconstruction module.

[0013] Furthermore, the anomaly detection model consists of a trained context information fusion module, a fuzzy embedded data generation module, a multiple spatiotemporal relationship explicit separation and representation module, and a reconstruction module; specifically, judging whether the gas turbine equipment is abnormal based on the difference between the process data reconstruction result at the last moment output by the anomaly detection model and the true value is as follows:

[0014] The mean square error of the process data reconstruction result at the last moment output by the calculation reconstruction module and the multiple spatiotemporal relationship explicit separation module is used as the anomaly score value.

[0015] If the abnormal score is greater than the abnormal threshold, it indicates that the gas turbine is operating in an abnormal state; otherwise, it indicates that the gas turbine is operating in a normal state.

[0016] Furthermore, the loss function is:

[0017]

[0018] Where λ1 and λ2 represent hyperparameters, and a represents the harmonic term. This indicates the reconstruction error of the reconstruction module. This represents the reconstruction error of the explicit separation and characterization module for multiple spatiotemporal relationships.

[0019] Furthermore, the membership function is a Gaussian function.

[0020] Furthermore, the output of the L-layer fully connected layer is represented as follows:

[0021]

[0022] In the formula, A t:t+1 This represents the normalized spatiotemporal graph, where σ represents the activation function. This represents the trainable weight parameters of the l-th fully connected layer. This represents the trainable bias term of the l-th fully connected layer.

[0023] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a new research approach for anomaly detection in the operating status of typical gas turbine equipment. By employing an anomaly detection model, which designs a customized spatiotemporal graph to perceive the dynamics of temporal dependencies, the continuity and lag of spatial coupling relationships in real time, it achieves effective separation and explicit extraction of multiple spatiotemporal relationships, thereby obtaining fine-grained temporal and spatial dependencies, thus providing a foundation for further analysis. Furthermore, a fuzzy embedding method based on fuzzy logic knowledge is designed, which for the first time cleverly maps process data into a high-dimensional fuzzy space to endow the process data with multiple fuzzy states. All fuzzy states are combined to enhance the characterization of multiple spatiotemporal relationships and solve their uncertainty problems, improving the interpretability and accuracy of anomaly detection in the operating process. The proposed method has been successfully applied in detailed experimental studies on typical gas turbine equipment. Ultimately, this method can be applied to the combined cycle power generation site of gas turbine equipment to ensure the safety and reliability of the gas turbine equipment operation process. Attached Figure Description

[0024] The invention will be further described below with reference to the accompanying drawings and embodiments:

[0025] Figure 1 This is a flowchart of the multi-fuzzy characterization and anomaly detection method for the spatiotemporal dynamics of gas turbine equipment described in this invention;

[0026] Figure 2 This is a schematic diagram of the fuzzy embedding method in this invention;

[0027] Figure 3 This is a schematic diagram of the reconstruction method based on fuzzy reasoning in this invention; Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0029] This invention uses a gas turbine as an example of a typical gas turbine device to further illustrate the invention. As a typical gas turbine device, the characteristics of a gas turbine are extremely complex, and its process data exhibits time-varying, dynamic, and non-stationary characteristics. Heavy-duty gas turbines are used for power generation and consist of three core components: a compressor, a combustion chamber, and a gas turbine. The compressor and turbine both employ a multi-stage axial flow design; the compressor has 15 stages with a pressure ratio of 16.9, and the turbine has 4 stages. The combustion chamber adopts a ring-shaped structure, arranged counterclockwise, and has an internal ceramic heat shield. The entire unit and fuel control valve assembly are located within the gas turbine casing. Its working principle is as follows: the compressor draws in air from the outside, which enters through the gas turbine inlet and is pressurized by the multi-stage blades of the compressor before being compressed and sent to the combustion chamber. Simultaneously, fuel (gaseous or liquid fuel) is also injected into the combustion chamber and mixed with the high-temperature compressed air for combustion at constant pressure. The high-temperature, high-pressure flue gas generated by combustion expands after being heated and enters the turbine zone. It is driven by multiple stages of blades to rotate the power blades at high speed until it is discharged from the outlet as exhaust gas. The exhaust gas can be discharged into the atmosphere or reused (such as by using a waste heat boiler for combined cycle). Due to changes in unit load, the fuel supply is constantly adjusted according to demand, resulting in continuous changes in the operating state of the gas turbine, forming a variety of different operating conditions.

[0030] This invention constructs a customized spatiotemporal graph to perceive the dynamics of temporal dependencies and the continuity and lag of spatial coupling relationships in real time, thereby achieving effective separation and explicit extraction of multiple spatiotemporal relationships. This results in fine-grained temporal and spatial dependencies, aiding in accurate anomaly detection. Furthermore, the frequent switching of operating conditions during gas turbine operation leads to uncertainties in multiple spatiotemporal relationships. A fuzzy embedding method is designed to map process data to a high-dimensional fuzzy space, endowing it with multiple fuzzy states, and effectively addressing the uncertainty problem through the joint representation of these fuzzy states. This invention not only adaptively achieves explicit separation and representation of multiple spatiotemporal relationships using a data-driven approach but also utilizes fuzzy logic knowledge to embed process data into fuzzy states, enhancing the characterization of multiple spatiotemporal relationships and resolving their uncertainty problems, thereby improving the interpretability and accuracy of anomaly detection. This invention provides a novel analytical perspective for the anomaly detection of typical gas turbine equipment operating status. It not only solves the shortcomings of existing methods that confuse time-dependent and spatial coupling relationships, but also enhances the ability to characterize multiple spatiotemporal relationships through the joint representation of fuzzy states. This improves the interpretability and accuracy of anomaly detection, helping engineers to make accurate judgments on the operating status of gas turbine equipment, detect abnormalities in a timely manner, and thus ensure the safety of actual production of gas turbine equipment.

[0031] Figure 1The flowchart shown is a process for the multi-fuzzy representation and anomaly detection method for the spatiotemporal dynamics of gas turbine equipment according to the present invention. It involves offline training and construction of an anomaly detection model, which is then used for online anomaly detection: real-time acquisition of process data fragments from the gas turbine equipment's operation is input into the trained anomaly detection model. The anomaly status of the gas turbine equipment is determined based on the reconstruction result of the process data at the last moment output by the anomaly detection model and the anomaly status of the true value. The offline training and construction of the anomaly detection model includes the following steps:

[0032] Step 1: Collect process data under frequently changing operating conditions of the gas turbine, and construct a training dataset by performing sliding window partitioning; this step is implemented by the following sub-steps:

[0033] (1.1) Collect normal operation data to be analyzed and obtain the operation state matrix: During the frequent change of operating conditions of the gas turbine system (frequent switching of operating conditions leads to frequent changes in data distribution), acquire process data collected at M time points during the normal continuous operation of the system, and the process data x' collected at each time point. t Each contains data from N measured variables, ultimately forming a two-dimensional state operation matrix.

[0034] In this embodiment, approximately 2000 process variable data points during normal operation were collected from a power plant in Zhejiang Province to construct an anomaly detection model. The measured variables are the following 32: pre-module natural gas volume flow rate, pre-module natural gas mass flow rate, compressor inlet temperature (representing ambient temperature), compressor outlet pressure 1, compressor outlet pressure 2, compressor outlet temperature, compressor bearing temperature 1, compressor bearing temperature 2, compressor bearing temperature 3, compressor thrust bearing generator end temperature 1, compressor thrust bearing generator end temperature 2, compressor thrust bearing generator end temperature 3, compressor thrust bearing gas turbine end temperature 1 1. Compressor thrust bearing gas turbine end temperature 2. Compressor thrust bearing gas turbine end temperature 3. Compressor bearing vibration 1. Compressor bearing vibration 2. Compressor side main shaft vibration. Compressor intake pressure difference. Gas turbine exhaust average temperature. Gas turbine side main shaft vibration. Combustion chamber pressure difference. Gas turbine cooling air regulating valve position 1. Gas turbine cooling air regulating valve position 2. Gas turbine humming 1. Gas turbine humming 2. Gas turbine speed. Gas turbine power. Gas turbine stage 2 stationary vane ring chamber cooling air pressure 1. Gas turbine stage 2 stationary vane ring chamber cooling air pressure 2. Gas turbine stage 3 stationary vane ring chamber cooling air pressure 1. Gas turbine stage 3 stationary vane ring chamber cooling air pressure 2.

[0035] (1.2) Perform a sliding window partitioning operation (HW) on the running state matrix to obtain the time-series extended data:

[0036] X' 1:T ,X' 2:T+1 ,...,X't-T+1:t ,...,X' M-T+1:M =HW(X')

[0037] in A sample represents time series data of N measured variables at T time points, where T represents the length of the sliding window.

[0038] Step 2: Construct a training network, which includes a context information fusion module, a fuzzy embedding data generation module, and a multiple spatiotemporal relationship explicit separation and representation module;

[0039] (2.1) First, in order to ensure that the fuzzy embedding process is not an isolated single-point embedding method, the context information fusion module needs to integrate the context information of the input sample along the time dimension into each variable based on the one-dimensional convolution kernel to obtain the context information fused sample, as shown below:

[0040] X t-T+1:t =X' t-T+1:t *θ ct

[0041] in X represents the context information convolution kernel, and δ represents the temporal span of context integration. t-T+1:t This represents the sample after obtaining contextual information fusion. Specifically, padding is required during convolution to ensure X' t-T+1:t and X t-T+1:t Size consistency between them.

[0042] (2.2) The fuzzy embedded data generation module includes a fuzzy membership value generation layer, a fuzzy membership vector generation layer, and a fuzzy embedded data generation layer, used for fuzzy embedded data generation (mapping data points to a high-dimensional space to give them multiple fuzzy states): Specific steps are as follows... Figure 2 As shown, this step is implemented by the following sub-steps:

[0043] (2.2.1) Constructing fuzzy sets: Considering that different variables have different fuzzy states, a specific fuzzy set MF is constructed for each variable. n (1≤n≤N), each fuzzy set contains Q distinct fuzzy membership functions.

[0044] (2.2.2) Fuzzy Membership Value Generation Layer generates fuzzy membership values: The data of each measured variable at each time step in the sample after context information fusion is input into the corresponding fuzzy set (containing Q membership functions; in this embodiment, the membership functions are Gaussian functions) to obtain Q fuzzy membership values. Taking the first sample X... 1:T The value of the nth measured variable at time t For example, it can be represented as:

[0045]

[0046] in This represents the q-th fuzzy membership value of the nth measured variable at time t. Represents the fuzzy membership function. This indicates the center of the membership function. This represents the bandwidth of the membership function. Specifically, and Adaptive learning can be performed during model training.

[0047] (2.2.3) The fuzzy membership vector generation layer obtains the fuzzy membership vector: it concatenates the outputs of Q fuzzy membership functions (Q fuzzy membership values) to generate the fuzzy membership vector corresponding to each measured variable at each time step.

[0048]

[0049] Where || represents the concatenation operation.

[0050] Then, a fuzzy membership vector with stronger representational power is obtained by using a nonlinear mapping function:

[0051]

[0052] in This represents the final fuzzy membership vector, which can effectively characterize... It has Q fuzzy states, Indicates the mapping weight parameters. This represents the mapping bias term.

[0053] (2.2.4) The fuzzy embedded data generation layer concatenates the fuzzy membership vectors of N measured variables at T time points to obtain fuzzy embedded data. Specifically, it first concatenates the fuzzy membership vectors corresponding to the values ​​of the nth measured variable at T time points to obtain... Corresponding fuzzy embedded data

[0054]

[0055] Then, the fuzzy embedded data corresponding to N measured variables are concatenated to obtain the final information about the sliding window segment. Fuzzy embedded data It can effectively characterize steady-state and transient operating conditions:

[0056]

[0057] (2.3) To achieve explicit separation of multiple spatiotemporal relationships, a customized cross-spatiotemporal graph needs to be designed to perceive temporal dependencies, temporal spatial coupling relationships, and cross-temporal spatial coupling relationships in real time. A multi-level spatiotemporal co-convolutional network is then used to mine deeper spatiotemporal dependency representations to obtain future prediction results. Specifically, a module for explicit separation and representation of multiple spatiotemporal relationships is used, which includes a cross-spatiotemporal graph generation module, a multi-level spatiotemporal co-convolutional network, and a prediction layer.

[0058] (2.3.1) The spatiotemporal graph generation module generates T-1 spatiotemporal graphs based on fuzzy embedded data: Specifically, it uses the previously generated fuzzy embedded data to calculate the correlation between all variables within two adjacent time points. Then, all correlations are aggregated to explicitly represent multiple spatiotemporal relationships:

[0059] (2.3.1.1) Divide the data into segments: Obtain the fuzzy embedded data FEX 1:T sub-fragments Furthermore, it is divided into smaller T-1 sub-segments every two moments along the leading edge time dimension. And reset the size of the sub-fragment to

[0060] (2.3.1.2) The T-1 spatiotemporal graphs and the T-1 fuzzy embedded data are recombined into sub-segments at every two time points along the leading edge time dimension, corresponding one-to-one. The spatiotemporal graph is a 2N×2N matrix, and the spatiotemporal graph CSG... t:t+1 The value in the i-th row and j-th column of (i,j) represents the correlation between the vectors in the i-th and j-th rows of the corresponding sub-segment:

[0061] CSG t:t+1 (i,j)=Cossim(SF t:t+1 (i),SF t:t+1 (j))

[0062] i,j=1,2,...,2N

[0063] Where i and j represent the i-th and j-th nodes in the spatiotemporal graph, respectively, SF t:t+1 (i) represents the sub-fragment SF t:t+1 The vector in the i-th row. Cossim(*) represents the cosine similarity;

[0064] (2.3.2) The multi-level spatiotemporal collaborative convolutional network contains T-1 spatiotemporal nodes. Figure 1A corresponding T-1 level, each level contains L fully connected layers and a max pooling layer connected sequentially. The input of the first fully connected layer in each level is the corresponding normalized spatiotemporal graph and the sub-segment. The input of the 2nd to Lth fully connected layers is the corresponding normalized spatiotemporal graph and the output of the previous fully connected layer. The output of the Lth fully connected layer is represented as follows:

[0065]

[0066]

[0067] in A represents the degree matrix of a spatiotemporal graph. t:t+1 This represents the normalized spatiotemporal graph, where σ represents the activation function. This represents the output of the l-th fully connected layer. This represents the trainable weight parameters. This represents a trainable bias term.

[0068] The max pooling layer aggregates and max pools the outputs of each fully connected layer:

[0069]

[0070]

[0071] Where Agg represents the aggregation operation. express The features of time points t and t+1 after time segmentation. Maxpooling represents maximum pooling.

[0072] (2.3.3) Finally, the prediction layer uses the T-1 level output concatenation result of the multi-level spatiotemporal co-convolutional network to predict and obtain the process data reconstruction result of the last moment of the sample.

[0073] U all =HM 1:2 HM 2:3 ||...||HM T-2:T-1

[0074]

[0075] Where W3 and W4 represent trainable weight parameters, and b3 and b4 represent trainable bias terms.

[0076] Step 3: Input each sample from Step 1 into the training network and train by minimizing the loss function until the loss function converges or the set number of training iterations is reached; the loss function includes the reconstruction error of the explicit separation and representation module of multiple spatiotemporal relationships, as shown below:

[0077]

[0078] in x represents the reconstruction error constraint of the explicit separation and representation module of multiple spatiotemporal relationships, and ||·||2 represents the L2 norm. T This represents the true value at time point T. This represents the predicted value at time point T, and N represents the number of variables.

[0079] The trained context information fusion module, fuzzy embedded data generation module, and multiple spatiotemporal relationship explicit separation and representation module can constitute an anomaly detection model for anomaly detection.

[0080] As a preferred embodiment, the training network of the present invention further includes a reconstruction module. A reconstruction method based on fuzzy inference reconstructs the samples, and the reconstruction error of the reconstruction module further constrains the training of the training network, improving the accuracy of the final anomaly detection model. Specifically, the reconstruction module includes a fuzzy rule layer, a fuzzy conclusion layer, and a defuzzification layer, with the steps as follows: Figure 3 As shown, it includes the following sub-steps:

[0081] (1) Fuzzy Rule Layer: This layer treats each neuron as a fuzzy rule and uses the T-norm operator to calculate the triggering intensity of the corresponding fuzzy rule:

[0082]

[0083] in This represents the triggering strength of the q-th fuzzy rule (the q-th neuron). This represents the fuzzy membership vector output by the q-th fuzzy membership function on the n-th variable.

[0084] (2) Fuzzy Conclusion Layer: This layer normalizes the trigger strength of all fuzzy rules and connects them together:

[0085]

[0086] FRN all =FRN(1)||FRN(2)||...||FRN(Q)

[0087] in This represents the normalized trigger strength. This represents the final trigger strength matrix.

[0088] (3) Deblurring layer: This layer uses nonlinear mapping to perform deblurring operations, generating reconstructed sliding window segments, i.e., samples.

[0089]

[0090] in Represents the reconstructed sample. Represents the weight parameters. This indicates the bias term.

[0091] During training, the reconstruction error of the reconstruction module is expressed as:

[0092]

[0093] in X represents the reconstruction error of the reconstruction module, and ||·||2 represents the L2 norm. 1:T This represents the original sample.

[0094] Furthermore, the evolutionary loss function can be jointly evolved in the following way to dynamically and collaboratively optimize all modules and comprehensively guide end-to-end training to complete the training:

[0095]

[0096] Where λ1 and λ2 represent hyperparameters, and a represents the harmonic term.

[0097] In this scheme, the anomaly detection model consists of a trained context information fusion module, a fuzzy embedded data generation module, and a reconstruction module, or it consists of a trained context information fusion module, a fuzzy embedded data generation module, a multiple spatiotemporal relationship explicit separation and representation module, and a reconstruction module. Specifically, for the anomaly detection model containing either a multiple spatiotemporal relationship explicit separation and representation module or a reconstruction module, the online anomaly detection steps are as follows:

[0098] (1) Obtain online test sliding window fragment Input it into the anomaly detection model to obtain the reconstruction result at the last time point.

[0099] (2) Calculate the anomaly score:

[0100]

[0101] Where ||·||2 represents the L2 norm, This indicates an abnormal rating value.

[0102] (3) Anomaly detection:

[0103]

[0104] Where η represents the pre-set abnormal threshold. This indicates that the gas turbine is operating under normal conditions. This indicates that the gas turbine has encountered an abnormality and requires further inspection and maintenance.

[0105] For an anomaly detection model that simultaneously includes explicit separation and reconstruction modules of multiple spatiotemporal relationships, the specific steps for online anomaly detection are as follows:

[0106] (1) Obtain online test sliding window fragment Inputting it into the anomaly detection model yields the reconstruction result output by the reconstruction module at the last time point. Reconstruction results output by the explicit separation module of multiple spatiotemporal relationships

[0107] (2) Calculate the anomaly score:

[0108]

[0109] Where ||·||2 represents the L2 norm, This indicates an abnormal rating value.

[0110] (3) Anomaly detection:

[0111]

[0112] Where η represents the pre-set abnormal threshold. This indicates that the gas turbine is operating under normal conditions. This indicates that the gas turbine has encountered an abnormality and requires further inspection and maintenance.

[0113] The effectiveness of the present invention will be further illustrated below with reference to the detection results of specific embodiments. The hyperparameters specified in the proposed method of the present invention are determined when the anomaly detection performance reaches its optimal level. Specifically, the number of fuzzy membership functions Q is set to 150, the sliding window length T is set to 10, λ1, λ2, and λ3 are set to 0.01, 0.9, and 0, respectively, and a is set to 2. Under the above hyperparameter settings, the anomaly detection model of the present invention is trained. The anomaly detection model consists of a trained context information fusion module, a fuzzy embedded data generation module, a multiple spatiotemporal relationship explicit separation and representation module, and a reconstruction module.

[0114] The anomaly detection model trained according to this invention was used to perform online real-time detection of a gas turbine operation process (including anomalies) containing 2000 samples. Table 1 shows the comparison results between the anomaly detection method of this invention and seven existing high-performance anomaly detection methods. As can be seen from Table 1, the method proposed in this invention exhibits the best performance (F1 score and AUROC index) in the gas turbine operation process, with its anomaly detection performance improved by 13.64% compared to the DAGMM method. The method proposed in this invention has higher sensitivity and robustness to changes in the gas turbine operating state, making it more suitable for practical applications of gas turbines. It helps industrial engineers make accurate judgments on the process operating state of gas turbine equipment, ensuring the safe and reliable operation of actual production processes.

[0115] Table 1 compares the detection results of the method of the present invention with those of seven existing anomaly detection methods.

[0116]

[0117] The proposed method for multi-fuzzy characterization and anomaly detection of gas turbine equipment, addressing the spatiotemporal dynamics of gas turbine systems, considers the multiple spatiotemporal relationships (temporal dependencies within subsystems, simultaneous spatial coupling between subsystems, and cross-timetime spatial coupling between subsystems) existing within subsystems during operation. It employs an anomaly detection model that constructs a customized cross-spatiotemporal graph to perceive the dynamics of temporal dependencies and the continuity and lag of spatial coupling in real time. This enables effective separation and explicit extraction of multiple spatiotemporal relationships, yielding fine-grained temporal and spatial dependencies. Furthermore, by introducing fuzzy logic knowledge, process data is mapped to a high-dimensional fuzzy space, endowing it with multiple fuzzy states. These fuzzy states are then combined to enhance the characterization of multiple spatiotemporal relationships and address uncertainty, ultimately improving the interpretability and accuracy of anomaly detection. Application to actual industrial processes has successfully demonstrated that this invention can provide timely and accurate anomaly detection results for technical management departments at actual gas turbine power generation sites, helping industrial engineers to take effective countermeasures and ultimately laying the foundation for the safe and reliable operation of industrial processes.

[0118] This invention is not limited to the operation process of the gas turbine equipment described in the specific embodiments above. Anyone skilled in the art can make equivalent modifications or substitutions without departing from this invention, and all such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for multi-fuzzy characterization and anomaly detection of the spatiotemporal dynamics of gas turbine equipment, characterized in that, Specifically, the process involves: acquiring real-time process data segments of the gas turbine equipment and inputting them into a trained anomaly detection model; then determining whether the gas turbine equipment is abnormal based on the difference between the reconstructed process data at the last moment output by the anomaly detection model and the true value; wherein, the anomaly detection model is trained through the following steps: Step 1: Collect process data under the frequent changes in operating conditions of the gas turbine during normal operation, and construct samples by sliding window partitioning to obtain a training dataset; where each sample is time series data of T times for N measured variables; Step 2: Construct a training network, which includes a context information fusion module, a fuzzy embedding data generation module, and a multi-spatiotemporal explicit separation and representation module. The context information fusion module integrates the context information of the input sample along the time dimension into each measured variable based on a one-dimensional convolutional kernel to obtain a context-fused sample. The fuzzy embedding data generation module includes a fuzzy membership value generation layer, a fuzzy membership vector generation layer, and a fuzzy embedding data generation layer. The fuzzy membership value generation layer generates Q fuzzy membership values ​​for each time step and each measured variable in the context-fused sample based on Q membership functions. The fuzzy membership vector generation layer concatenates and enhances the Q fuzzy membership values ​​corresponding to each time step and each measured variable to obtain a fuzzy membership vector. The fuzzy embedding data generation layer concatenates the fuzzy membership vectors of T time steps and N measured variables to obtain fuzzy embedding data. The multi-spatiotemporal explicit separation and representation module includes a cross-spatiotemporal graph. The system comprises a generation module, a multi-level spatiotemporal co-convolutional network, and a prediction layer. The spatiotemporal graph generation module generates T-1 spatiotemporal graphs based on fuzzy embedded data. Each of the T-1 spatiotemporal graphs corresponds one-to-one with a sub-segment of the T-1 fuzzy embedded data, recombined at every two time points along the leading edge time dimension. The spatiotemporal graph is a 2N×2N matrix, where the value in the i-th row and j-th column of the spatiotemporal graph represents the correlation between the i-th and j-th row vectors of the corresponding sub-segment. The multi-level spatiotemporal co-convolutional network contains T-1 levels corresponding one-to-one with the T-1 spatiotemporal graphs. Each level consists of L fully connected layers and a max pooling layer connected sequentially. The input of the first fully connected layer in each level is the corresponding normalized spatiotemporal graph and sub-segments. The input of the 2nd to Lth fully connected layers is the corresponding normalized spatiotemporal graph and the output of the previous fully connected layer. The max pooling layer is used to aggregate the outputs of each fully connected layer and perform max pooling. The prediction layer is used to predict the process data reconstruction result of the last moment of the sample based on the concatenation result of the output of the T-1 level of the multi-level spatiotemporal co-convolutional network. Step 3: Input each sample from Step 1 into the training network and train it by minimizing the loss function until the loss function converges or the set number of training iterations is reached; the loss function includes the reconstruction error of the multiple spatiotemporal relationship explicit separation and representation module; the anomaly detection model consists of a trained context information fusion module, a fuzzy embedding data generation module, a reconstruction module, and a multiple spatiotemporal relationship explicit separation and representation module.

2. The method according to claim 1, characterized in that, The training network further includes a reconstruction module, which comprises a fuzzy rule layer, a fuzzy conclusion layer, and a defuzzification layer. The fuzzy rule layer contains Q neurons, and each neuron uses the T-norm operator to calculate the trigger strength of the corresponding fuzzy rule based on the fuzzy membership vector output by the fuzzy membership function of the nth variable in the fuzzy embedded data. The fuzzy conclusion layer is used to normalize and concatenate the trigger strengths output by the fuzzy rule layer. The deblurring layer is used to deblur the output of the fuzzy conclusion layer using a nonlinear mapping to generate reconstructed samples; the loss function also includes the reconstruction error of the reconstruction module.

3. The method according to claim 2, characterized in that, The anomaly detection model consists of a trained context information fusion module, a fuzzy embedded data generation module, and a reconstruction module, or it consists of a trained context information fusion module, a fuzzy embedded data generation module, a multiple spatiotemporal relationship explicit separation and representation module, and a reconstruction module.

4. The method according to claim 2, characterized in that, The anomaly detection model consists of a trained context information fusion module, a fuzzy embedded data generation module, a multiple spatiotemporal relationship explicit separation and representation module, and a reconstruction module. Specifically, determining whether the gas turbine equipment is abnormal based on the difference between the reconstructed process data at the last moment output by the anomaly detection model and the true value involves: The mean square error of the process data reconstruction result at the last moment output by the calculation reconstruction module and the multiple spatiotemporal relationship explicit separation module is used as the anomaly score value. If the abnormal score is greater than the abnormal threshold, it indicates that the gas turbine is operating in an abnormal state; otherwise, it indicates that the gas turbine is operating in a normal state.

5. The method according to claim 2, characterized in that, The loss function is: Where λ1 and λ2 represent hyperparameters, and a represents the harmonic term. This indicates the reconstruction error of the reconstruction module. This represents the reconstruction error of the explicit separation and characterization module for multiple spatiotemporal relationships.

6. The method according to claim 1, characterized in that, The membership function is a Gaussian function.

7. The method according to claim 1, characterized in that, The output of the L-layer fully connected layer is represented as follows: In the formula, A t:t+1 This represents the normalized spatiotemporal graph, where σ represents the activation function, and W1 l W2 l This represents the trainable weight parameters of the l-th fully connected layer. This represents the trainable bias term of the l-th fully connected layer.

Images